SYSTEM ARCHITECTURE
User
↓
Hermes Orchestrator (8039)
↓ Task Classifier
Gemini Flash / GPT-5 nano
↓ Routes to
Coding → GPT-5.5 / Claude Opus
Reasoning → o3 / GPT-5.5
Vision → Qwen-VL / Gemini Vision
Local → Ollama (Qwen3, Gemma, DeepSeek)
Fast → Gemini Flash
↓ Provider Layer
LiteLLM Proxy
9Router (fallback)
↓ Providers
OpenAI
Anthropic
OpenRouter
Ollama
Azure
vLLM
The runtime platform: agent execution engine, DAG orchestration, event bus, model routing (local/cloud), and the core API surface. Everything in AgentsOS runs on top of this.
Status✅ Shipped v3.1 — 80+ tests, production-ready
ComponentsControl Plane (8080), DAG Engine, Event Bus, PDO Models
StackPython FastAPI, Pydantic, SQLAlchemy, pytest, ruff
NextPostgreSQL migration, JWT auth, durable task queue (v3.2)
Takes a project brief, plans it as a DAG, resources it against the agent fleet, executes it end-to-end, reviews results, and learns from every outcome. The core intelligence of the platform.
LifecycleBrief → PDO → Task Plan → DAG → Execute → Review → Learn
Hermes EngineDAG planning & scheduling (port 8039)
Safety ModesYOLO, Review, Approval — per-task gating
Coming in v4.0LLM-native hierarchical planner, evolution wiring, provenance-gated execution
Six specialist agents, each with its own modality, load tracking, and concurrency model. The fleet scales per-core with auto-tuned headroom.
🖥️ Code AgentCode generation, editing, refactoring
☁️ Infra AgentDocker, Cloudflare, Tailscale, cloud console
👁️ Vision AgentScreenshot analysis, diagram understanding, OCR
🌐 Browser AgentPlaywright-based web automation (Phase 2)
📱 Android AgentDevice control via ADB + Hermes Bridge (Phase 2–3)
📄 Docs AgentRAG-powered document read/write/summary
Your conversational project director. Desky lives in the MyDesk dashboard, speaks DAG, generates briefs, and guides you through project lifecycle from kickoff to lessons learnt.
CapabilitiesDAG preview, approval gates, multi-turn planning
IntegrationMyDesk frontend persona — cosmetic layer, no infra
ModesChat, Brief Generator, Project Review, Lessons Learnt
ComingDesky Pro — multi-project portfolio view (v4.1)
The step executor and memory graph. OpenClaw runs individual task nodes using a fixed set of primitives (run_command, read/write, http_request, model_infer), writes every action to the provenance ledger, and validates results against acceptance criteria.
Port4096
LedgerEvery action writes to Neo4j provenance graph
Primitivesrun_command, read/write, http_request, model_infer, run_python, run_node
Coming in v2Sandboxed execution, rollbacks, retry policies (v4.0)
Three-layer routing architecture: Agent Router (which specialist) → Model Router (which LLM) → Provider Router (which endpoint). Task-classifier-driven with automatic fallback chains.
Layer 1 — AgentHermes classifies intent → routes to Code/Infra/Vision/Browser/Android/Docs agent
Layer 2 — ModelTask type determines model: coding→GPT-5.5, reasoning→o3, vision→Qwen-VL, fast→Gemini Flash
Layer 3 — ProviderLiteLLM → 9Router fallback → cheapest/healthiest endpoint
9Router122 models, 40+ providers, automatic tiered fallback (subscription → cheap API → free). Available as Docker profile.
A registry of execution tools and skills that agents can discover and invoke. MCP-compatible connectors expand the platform's reach into any external system.
Port7070
ConnectorsGitHub, Cloudflare, Tailscale, Docker, Slack, Teams
DiscoveryAgents query registry for available tools by capability tag
ExtensibleAny MCP-compatible server can register itself
The knowledge and memory layer. Qdrant provides semantic/vector retrieval (RAG) over docs and code. Neo4j stores the agent activity ledger and capability/relationship graph. Redis handles short-term session memory and task queues.
Qdrant (6333)Vector store — RAG over code, docs, customer data
Neo4j (7474/7687)Graph store — provenance ledger, agent relationships
Redis (6379)Queue, cache, session memory
ComingDigital Twin — personalisation layer over ledger (Phase 3)
The human-facing workspace: agent dashboard, chat, DAG visualisation, and business verticals (legal, finance, CRM). Consumes AgentsOS over the Control Plane API exclusively — never talks to agents directly.
StackBlazor / MudBlazor 7.15, .NET 10, Next.js dashboard shell
VerticalsLegal (AU Consumer Law, QLD Legal Practice Act) — pilot ready
ControlsProject dashboard, DAG visualisation (D3.js), approval gates
IntegrationREST + WebSocket to AgentsOS API. Never owns orchestration logic.
First-party AI orchestration agent. Hermes handles model switching via hermes model, provider abstraction (OpenAI/Anthropic/Ollama/vLLM/OpenRouter), gateway routing, and serves as the DAG planning engine inside Master Orchestrator.
Port8039
Auto-RoutingIntent classifier → role-based session routing (RFC)
FallbackAutomatic provider failover, credential pools
IntegrationBridged with AgentsOS Docker stack. Windows pb-legacy ↔ Docker.
ℹ️ Hermes Routing Architecture — The platform uses a three-layer routing model: Agent Router (which specialist agent) → Model Router (which LLM) → Provider Router (which endpoint). This avoids combining all decisions into one monolithic router and follows the industry pattern emerging across 9Router, LiteLLM, and the Hermes Auto-Routing RFC.
Optional AI gateway for provider-level failover and cost optimisation. 9Router sits behind LiteLLM as a fallback layer, managing subscription tiers, API quotas, and automatic failover across 40+ providers.
Status✅ Running on localhost:20128 — 122 models
Docker Profiledocker compose --profile 9router up -d
Skills8 Hermes skills installed (chat, image, tts, stt, embeddings, web-search, web-fetch)
Tiered FallbackSubscription → Cheap API → Free providers, no user interaction